Back to all steps
Ask AI
Step fields
1
selector
(text)
2
question for AI (optional)
(text)
3
slot
(slot)
4
webhook endpoint
(webhook)
In a scenario, fields are joined with
|.Ask the model a question about the page content. Format: selector|instruction|slot|webhook.
Example: .product-card|Extract the price as a number|slot_1| plus an endpoint.
The selector may be omitted - then the model reads the visible text of the whole page (up to 12,000 characters).
The answer goes into a slot and is then substituted into any step as rmb:slot_1 - so a single step covers both data extraction and generating text to fill a form.
The step fails when the model cannot fulfil the instruction from the page text: the data is missing, or the answer to the question is no. So the instruction can be written as a check: «is the price visible and non-zero?».
Unlike «Page perception», which looks at the screenshot and judges the layout, this one reads the text and judges the content - at a fraction of the tokens. When tokens run out the step is skipped rather than failing the test.
Limits and safety. At most 12,000 characters of the page reach the model, the instruction is capped at 2,000, and the answer is truncated to 2,000 characters and stripped of control characters. The page text is passed to the model as untrusted data: directives found inside it must not be followed. No known technique makes this airtight, so treat the answer as data coming from the site under test: it lands in a slot and in a webhook, and on a third-party site its content can be influenced by that site's owner. For monitoring your own sites this does not matter.
Example: .product-card|Extract the price as a number|slot_1| plus an endpoint.
The selector may be omitted - then the model reads the visible text of the whole page (up to 12,000 characters).
The answer goes into a slot and is then substituted into any step as rmb:slot_1 - so a single step covers both data extraction and generating text to fill a form.
The step fails when the model cannot fulfil the instruction from the page text: the data is missing, or the answer to the question is no. So the instruction can be written as a check: «is the price visible and non-zero?».
Unlike «Page perception», which looks at the screenshot and judges the layout, this one reads the text and judges the content - at a fraction of the tokens. When tokens run out the step is skipped rather than failing the test.
Limits and safety. At most 12,000 characters of the page reach the model, the instruction is capped at 2,000, and the answer is truncated to 2,000 characters and stripped of control characters. The page text is passed to the model as untrusted data: directives found inside it must not be followed. No known technique makes this airtight, so treat the answer as data coming from the site under test: it lands in a slot and in a webhook, and on a third-party site its content can be influenced by that site's owner. For monitoring your own sites this does not matter.